Fault Detection and Diagnosis for Sensor in an Aero-Engine System

Zhen Zhao1, Yi-gang Sun1, Jun Zhang2

1. College of Electronics Information and Automation, 2. College of Aerospace Engineering, Civil Aviation University of China, Tianjin
300300, China
Email: zhenzhao0523@gmail.com

Abstract: In modern society, civil plane has become more and more important in people's life. How to ensure safety and
reliability of civil plane has attracted more and more attention by many research institutes and researchers. As one of the
most important part in the civil plane, how to ensure normal operation of aero-engine becomes particularly important.
According to statistics, the sensor fault accounts for more than 80% of the total failures in the aero-engine system. In this
paper, an aero-engine model is established according to the operational principle of aero-engine using AMESim
software. Secondly, the sensor sub-model in above model is modified to simulate drift failure, bias fault, and pulse fault,
and to achieve aero-engine sensors’ fault data. After that, sensor fault diagnosis model is built based on principle
component analysis. Online detection and diagnosis can be realized using this diagnosis model. Simulation results show
that the proposed method is feasible and effective.
Key Words: Sensor; Aero-engine; Fault detection and diagnosis; Principle component analysis

an analytical redundancy-based approach is presented for

1 INTRODUCTION
In modern society, civil aviation has occupied a particularly
important position in people’s learn, work, and life.
However, the ever increasing need also brings more and
more severe safety and reliability requirement, especially
after MH370 accident. Therefore, many research institutes
and scholars have paid more attention on how to ensure the
safety and reliability of the civil plane. As the heart of a civil
plane, safety and reliability of an aero-engine system is
particularly important for two reasons: one is that most of
plane accidents are caused by its aero-engine system, and
the other is that any tiny fault may well cause a major
aircraft accidents [1]. Furthermore, the operating
environment of the aero-engine is always very harsh, high
temperature, strong vibration, high EMI, high pressure, and
etc. Due to the harsh operating environment, the
aero-engine is very prone to failure. According to statistics
results supplied by relevant department, the sensor fault
accounts for more than 80% of the total faults in the
aero-engine system. Considering the abovementioned
reason, proposing effective fault detection and diagnosis
method for sensor in the aero-engine system will greatly
improve reliability and safety of the civil plane, and shows
very important theoretical and practical significance.
In the past two decades, lots of researches have been made
on sensor fault diagnosis for the aero-engine system. In
1987, analytical redundancy method has been proposed to
detect sensor’s bias and drift failures, whose performance
has been evaluated in a test-bed engine system [2]. In 1998,
This work is supported by the project 973 (Grant No. 2014CB744904),
Civil aviation science and technology major projects (Project No.
MHRD20130112), Joint Funds of the National Natural Science
Foundation of China and Civil Aviation Administration of China Key
Project (Grant No. U1533201), the Fundamental Research Funds for the
Central Universities (Project No. 3122014C010).
sensor fault detection of a single shaft industrial gas turbine
[3]. At the same year, a fuzzy logic based accommodation
approach for redundant sensor selection is proposed in [4].
In 1999, a sensor fault detection method using I/O linear
models to generate residuals permitting the identification of
possible sensor faults [5]. After 2000, types of pattern
recognition methods have been introduced to detect and
diagnose sensor faults, such as neural network [6-8].
Although QAR can record much sensor data for modern
aero-engine, the faulty sensor data is still very rare due to
that once a fault happens, some protection measure must be
implemented to make sure safety operation of the civil plane.
Considering few faulty sensor data, this paper firstly builds
an aero-engine model based on AMESim software, which
can realize simulation in both normal condition and
condition with sensor fault. Then, a statistical analysis
method is introduced to realize sensor fault detection and
diagnosis. Finally, the proposed statistical based sensor
fault detection and diagnosis method is experimented using
simulated sensors’ data achieved from the aero-engine
model.
The rest of this paper is organized as below: The
aero-engine model with modified sub-model for sensor is
depicted in Section 2. The sensor fault detection and
diagnosis method is described in Section 3. In Section 4, the
proposed fault detection and diagnosis method is illustrated
using simulated sensors’ output data. Finally, the
conclusions are drawn in Section 5.
2 AERO-ENGINE MODEL
With the rapid development of science and technology,
large quantity of operation data (including normal and
abnormal data) can be recorded through QAR in modern
civil plane. However, from a safety perspective, the whole

978-1-4673-9714-8/16/$31.00 2016
c IEEE 2977

Authorized licensed use limited to: East China Univ of Science and Tech. Downloaded on September 29,2022 at 08:39:58 UTC from IEEE Xplore. Restrictions apply.
 control system must take some protection measure
immediately after any fault. Therefore, faulty data are still
lack. To present a more feasible sensor fault diagnosis
method, this paper builds an aero-engine model using
AMESim software, which combines lots of module suitable
for dynamic characteristics’ simulation, and has been
proved to be effective in modeling and simulation of
engineering machinery hydraulic system.

2.1 Operation Principle of Aero-Engine Fig. 1 Structure of a turbofan aero-engine

Most of aero-engine in civil aviation is turbofan aero-engine, 2.2 Aero-Engine Model

which consists of fan, low-pressure compressor, According to the structure of a turbofan aero-engine, the
high-pressure compressor, combustion chamber, high whole model for the turbofan aero-engine in AMESim is
pressure turbine (driving compressor), low pressure turbine modeled as shown in Fig. 2.
(driving fan), exhaust system, etc., as shown in Fig. 1.

Fig. 2 AMESim model for a turbofan aero-engine

2.3 Sensor Model

In AMESim, it is not able to realize sensor fault simulation
through its plug-in module. Thus, it is needed to rebuild
appropriate sensor modules or modify the original sensor
modules. This paper chooses the second method.
For temperature sensors, firstly, drag the temperature sensor Fig. 4 Temperature sensor for high-pressure compressor
modules from pneumatic components design library, and For speed sensors, two speed sensors are dragged from the
connect them to low pressure compressor and high pressure machine design library, and connected to the low speed
compressor in Fig. 2, respectively. Then, add Add controller shaft and the high speed shaft as shown in Fig. 5 and 6,
and Multiply controller to the 2# output port of each respectively.
temperature sensor. The other input port of each controller
can be connected to relevant signal. Finally, sensor fault
data can be obtained through the 2# output port of each
controller.

Fig. 5 Speed sensor for low speed shaft

Fig. 3 Temperature sensor for low-pressure compressor

2978 2016 28th Chinese Control and Decision Conference (CCDC)

Authorized licensed use limited to: East China Univ of Science and Tech. Downloaded on September 29,2022 at 08:39:58 UTC from IEEE Xplore. Restrictions apply.
 where Ok (k 1, , m) is the contribution rate of each
principle component; Cl  [0,1] is threshold of
contribution, which is always set to be 95%.
This paper chooses SPE statistics as the monitoring
statistics. The SPE statistics can be calculated as bellow:
SPE ei eiT (4)
Fig. 6 Speed sensor for high speed shaft th
where ei is the i row of the residual matrix E , which
3 SENSOR FAULT DIAGNOSIS in an denotes the residual error at the i th sampling point.
2
AERO-ENGINE SYSTEM
When SPE exceeds its confidence limit QD staying for
2
A typical statistical analysis method, i.e. principle
some time, there may be sensor fault. QD can be set as
component analysis (PCA), has been widely applied in
below:
many research domain, such as statistical process control,
1/ h0
process monitoring, fault detection and diagnosis [1, 8], and ª CD 2T 2 h02 T 2 h0 (h0  1) º
so on. To now, PCA have achieved many accomplishments QD T1 « 1 » (5)
¬ T1 T12 ¼
in these fields. So this paper tries to detect and diagnose
sensor fault using PCA.
2.4 On-line Sensor Diagnosis
For fault diagnosis, PCA has two phases, i.e. off-line
modeling phase and on-line diagnosis phase. On the off-line To realize on-line sensor diagnosis, new sampled data xnew
modeling phase, the normal multi-variables’ data is can be decomposed using the built diagnosis model as
subjected to PCA to build a diagnosis model and get the below.
statistics of T 2 , SPE, and their corresponding control limits. xnew xˆnew  xnew (6)
On the on-line diagnosis phase, new coming
multi-variables’ data is reflected to the diagnosis model to where xnew PPT xnew ( I  C ) xnew Cxnew ,
get T 2 statistics or SPE statistics or both. According to the xˆnew T
PP xnew Cxnew , I is an identity matrix.
statistics and their corresponding control limits, diagnosis
result can be deduced. Then, the SPE statistics at a new sampling time can be
calculated as Eq. (7).
3.1 Sensor Diagnosis Model 2
SPE ( xnew ) xnew T
xnew ( I  C ) xnew
Given a normal modeling data matrix X c  R K uJ
composed (7)
T
xnew ( I  P PT ) xnew
of K samples of J variables (i.e. sensor outputs), which is
firstly normalized to a zero-mean and unit-variance matrix
4 ILLUSTATION AND DISCUSSION
X  R K u J using Eq. (1).
X [ X c  I m u T ]DV1/ 2 (1) In this section, the proposed statistics based fault detection
and diagnosis approach is applied to data obtained from the
where I m is a J u1 column vector whose elements are all 1, built aero-engine model. In the model, parameters are set as
u [u1 , , u J ]T is a mean vector of each variable in X c , default. However, simulation time and height is set
according to practical flight data. Simulation time is set as
DV diag (V 12 , , V n2 ) is a variance matrix of each variable
5400s, i.e. 1.5h. Sample period is set as 1s. Height is set as a
in X c . segmented function in AMESim as shown Fig. 7.
After that, PCA aims to decompose X into
X Xˆ  X TPT  TPT TPT  E (2)
where T  R K ul and P  R J ul are the score and loading
matrices, respectively; E is the residual matrix which can
be further decomposed into the product of a residual score
matrix T  R K u( J l ) and a residual loading matrix
P  R J u( J l ) ; l is the number of principal components (PCs)
retained in the statistical model, and is determined using Eq.
(3).
i

¦O k
k 1
m
t Cl (3)
¦ Oi
i 1

2016 28th Chinese Control and Decision Conference (CCDC) 2979

Authorized licensed use limited to: East China Univ of Science and Tech. Downloaded on September 29,2022 at 08:39:58 UTC from IEEE Xplore. Restrictions apply.

Fig. 7 Flight height setting in AMESim
Figs. 8 and 9 show the temperature, pressure, and speed
under normal condition in the aero-engine model. The
simulation results are consistent with practical data.
However, noise should be added to make the data be nearer
with practical data.
Fig. 8 Temperature and pressure in the aero-engine system
Fig. 9 Speed of the low speed shaft and high speed shaft
Then, four sensors’ output data, i.e. FF (Fuel Flow), EGT
(Exhaust Gas Temperature), N1 (Low Pressure Rotor
Speed), N2 (High Pressure Rotor Speed) and TAT (Total
Air Temperature), are recorded under normal and types of
fault conditions.
The sampled data is firstly normalized. Using the normal
data to build fault detection and diagnosis model, Table 1
2980 2016 28th Chinese Control and Decision Conference (CCDC)
Authorized licensed use limited to: East China Univ of Science and Tech. Downloaded on September 29,2022 at 08:39:58 UTC from IEEE Xplore. Restrictions apply.
shows the contribution rate at off-line modeling stage. The
first two principle components’ cumulative contribution of
variance is 96.63% in Table 1, which is larger than 95%. For
this reason, the number of retained PCs is 2, and the control
limits are set different at different stage.
Table1. Contribution rate
Number of Primary cumulative
Primary variance
principle contribution of variance
contribution rate (%)
components (%)
1 67.11 67.11
2 23.51 90.62
3 6.01 96.63
4 1.76 98.39
5 1.61 100
4.1 No Fault
Fig. 10 shows the SPE statistics under normal condition at
on-line stage. From Fig. 10, it can be seen that the SPE
statistics is almost under the corresponding control limit,
which illustrates that the sensors are all correct.
Time/s
Fig. 10 No fault for temperature sensor of high-pressure compressor
(dashed line - control limit, solid line - SPE statistics)
4.2 Drift Bias Fault
A drift bias of 0.005 is added into the temperature sensor for
high pressure compressor after the 2700th sampling point.
Fig. 11 shows the drift bias fault occurs to the temperature
sensor of the high pressure compressor
Fig. 11 Drift bias fault of temperature sensor for high pressure compressor
Fig. 12 illustrates the diagnosis result for this drift bias fault.
From Fig. 12 (a), it can be seen that the SPE statistics
exceeds the corresponding control limit at about 3681
sampling point. The reason may be that the drift bias is quite
little to be detected by the diagnosis model. Fig. 12 (b) and
(c) give the contribution rate histogram at 2800th sampling
point and 4000th sampling point, which demonstrates that
 the fifth variable (i.e. temperature sensor output of high
pressure compressor) shows largest contribute rate for this
phenomenon. This means the temperature sensor of high
pressure compressor occurs fault.
Time/s
(a) SPE statistics
Contribution rate
Variable
(b) Contribution rate histogram during cruise phase (at 2800)
Contribution rate
Variable
(c) Contribution rate histogram during landing phase (at 4000)
Fig. 12 Drift bias fault for temperature sensor of high-pressure compressor
4.3 Offset Fault
An offset of 0.05 is added into the temperature sensor of
high pressure compressor from the 2700th sampling point.
Fig. 13 shows the offset fault occurs to the temperature
sensor of high pressure compressor.
Fig. 13 Offset fault for temperature sensor of high-pressure compressor
Fig. 14 gives the detection and diagnosis results. In Fig. 14
(a), the SPE statistics exceeds its corresponding control
limit at about 2700th sampling point. Fig. 14 (b) and (c)
shows the contribution rate histogram at 2800th sampling
point and 4600th sampling point. From this figure, it
2016 28th Chinese Control and Decision Conference (CCDC)
Authorized licensed use limited to: East China Univ of Science and Tech. Downloaded on September 29,2022 at 08:39:58 UTC from IEEE Xplore. Restrictions apply.
demonstrates that the fifth variable (i.e. temperature sensor
output of high pressure compressor) shows largest
contribute rate for this phenomenon, which means that the
temperature sensor of high pressure compressor occurs
fault.
Time/s
(a) SPE statistics
Contribution rate
Variable
(b) Contribution rate histogram during cruise phase (at 2800)
Contribution rate
Variable
(c) Contribution rate histogram during landing phase (at 4600)
Fig. 14 Offset fault for temperature sensor of high-pressure compressor
4.4 Pulse Fault
Pulse fault is similar to offset fault. In this simulation, pulse
fault is added into the temperature sensor from the 2700th
sampling point. Fig. 15 shows the pulse fault occurs to the
temperature sensor of high pressure compressor.
Fig. 15 Pulse fault for temperature sensor of high-pressure compressor
Fig. 16 (a) shows the SPE statistics for pulse fault, Fig. 16 (b)
and (c) illustrates the contribution rate histogram at the
2800th and 4600th sampling point. From Fig. 16 (a), it can be
2981
 seen that the SPE statistics exceeds it corresponding control
limit at about the 3000th sampling point. Fig. 16 (b) and (c)
demonstrate that the fifth variable (i.e. temperature sensor
output of high pressure compressor) shows largest
contribute rate for this phenomenon. This means the
temperature sensor output of high pressure compressor
occurs fault. The detection and diagnosis result conforms to
the set.
Time/s
(a) SPE statistics (dash line – control limit, solid line – SPE statistics)
Contribution rate
Variable
(b) Contribution rate histogram during cruise phase (at 2800)
Contribution rate
Variable
(c) Contribution rate histogram during landing phase (at 4600)
Fig. 16 Pulse fault for temperature sensor of high-pressure compressor
5 CONCLUSION
In this paper, an aero-engine model is firstly modelled using
the AMESim software according to the principle of
aero-engine model. After that, a statistics based fault
detection and diagnosis method is introduced through
principle component analysis. On the diagnosis process,
2982 2016 28th Chinese Control and Decision Conference (CCDC)
Authorized licensed use limited to: East China Univ of Science and Tech. Downloaded on September 29,2022 at 08:39:58 UTC from IEEE Xplore. Restrictions apply.
normal data obtained from the built aero-engine model is
used to build the off-line diagnosis model. Some types of
sensor faults, such as drift bias fault, offset fault and pulse
fault, are used to test the detection and diagnosis model.
Simulation results show the feasibility and effectiveness of
the proposed detection and diagnosis method in the
aero-engine system.
REFERENCES
[1] Zhen Zhao, Yi-gang Sun, Jun Zhang, “PCA-Based Sensor
Fault Diagnosis for Aero-Engine,” 2015 CCDC, 2015,
2679-2683.
[2] W. C. Merrill, J. C. DeLaat, and W. M. Bruton, “Advanced
detection, isolation, and accommodation of sensor failures
real-time evaluation,” NASA, pp. 1–27, 1987.9.
[3] S. Simani, P. Spina, S. Beghelli, R. Bettocchi, and C.
Fantuzzi, “Fault detection and isolation based on
dynamic observers applied to gas turbine control
sensors,” ASME, 1998.
[4] T. Healy, L. Kerr, and L. Larkin, “Model based fuzzy
logic sensor fault accommodation,” Journal of
Engineering for Gas Turbines and Power, vol. 120, pp.
533–536, 1998.
[5] N. Aretakis, K. Mathioudakis, and A. Stamatis,
“Identification of sensor faults on turbofan engines
using pattern recognition techniques,” Control
Engineering Practice, vol. 12, pp. 827–836, 2004.
[6] M. Napolitano, A. Younghwan, and B. Seanor, “A fault
tolerant flight control system for sensor and actuator
failures using neural networks,” Aircraft Design, vol. 3,
pp. 103–128, 2000.
[7] K. Botros, G. Kibrya, and A. Glover, “A demonstration
of artificial neural networks based data mining for gas
turbine driven compressor stations,” ASME, vol. 3, pp.
2000–GT–0351, 2000.
[8] P. Dewallef and O. Leonard, “On-line validation of
measurements on jet engines using automatic learning
methods,” ISABE, pp. 2001–1031, 2001.
[9] Zhao Chunhui, Wang Fuli, Lu Ningyun, Jia Mingxing.
Stage-based Soft-transition Multiplepca Modeling and
On-line Monitoring Strategy for batch processes [J],
Journal of Process Control, 2007, 17(9): 728-741.
[10] Yue, H. Henry, Qin, S. Joe. Reconstruction-based fault
identification using a combined index [J], Industrial &
Engineering Chemistry Research, 2001, 40(20):
4403-4414.